Agent_Zero�and �Generative Social Science

Joshua M. Epstein, Ph.D.

New York University

​

Alphabet

October 12 2021

Affiliations

• Professor of Epidemiology NYU School of Global Health
• Director of the NYU ABM_Lab
• Affiliated Faculty at the Courant Institute of Mathematical Sciences and the Department of Politics, NYU.
• External Faculty Fellow. Santa Fe Institute

​

Generative explanation*

• To explain a social regularity
• Demonstrate how it could emerge on time scales of interest to humans in a population of cognitively plausible agents
• Does the micro-specification m generate the macroscopic explanandum x
• If so, m is a generative explanatory candidate.
• Motto (Epstein, 1999) is negative : If you didn’t grow it, you didn’t explain it.

​

​

• Not the converse (any old way of growing it is explanatory).
• Not uniqueness (might be many m’s).
• Generative sufficiency a necessary (but not sufficient) condition for explanation.
• ¬Furnish a Game in which the pattern is Nash
• ¬Furnish a Functional with respect to which the trajectory is an extremal
• ¬Furnish a Regression relating aggregate variables.

​

* … as against prediction.

​

Generative explanation*

• To explain a social regularity
• Demonstrate how it could emerge on time scales of interest to humans in a population of cognitively plausible agents
• Does the micro-specification m generate the macroscopic explanandum x
• If so, m is a generative explanatory candidate.
• Motto (Epstein, 1999) is negative : If you didn’t grow it, you didn’t explain it.

​

​

• Not the converse (any old way of growing it is explanatory).
• Not uniqueness (might be many m’s).
• Generative sufficiency a necessary (but not sufficient) condition for explanation.
• ¬Furnish a Game in which the pattern is Nash
• ¬Furnish a Functional with respect to which the trajectory is an extremal
• ¬Furnish a Regression relating aggregate variables.

​

* … as against prediction.

​

Cognitively Plausible Agents

• Have emotions (notably fear)
• Have bounded deliberative capacity
• Have social connection
• And all of those might matter.

​

​

​

​

​

Consistent Theme in Philosophy and Literature

Hume, “Reason is a slave to the passions.”

Aristotle, “Man is by nature a social animal.”

​

Looking for a simple convolution:

​

Passion ⊕ Reason ⊕ Social → Agent_Zero

​

Agent_Zero: Toward Neurocognitive

Foundations for Generative Social Science

Princeton University Press 2013

​

​

​

​

​

​

Funded by an NIH Director’s Pioneer Award

​

​

Agent_Zero

​

Endowed with distinct affective, deliberative, and social modules each grounded in contemporary neuroscience:

Internal modules interact to produce observable individual behavior.

Multiple agents interacting generate wide variety of collective dynamics: health, conflict, network dynamics, economics, social psychology, law.

Modules can be refined…Get synthesis started.

All provisional….

​

But Formal

Lots of empirical criticisms of the rational actor of Economics and Game Theory.

Gripes (even decisive experiments) do not change scientific practice.

Need explicit formal alternatives.

Albeit provisional, Agent_Zero is one: mathematical and computational.

​

​

The basic idea

• Before laying out the equations or showing code

Big Picture: The violence interpretation�

• Agents occupy an landscape of indigenous sites (yellow)
• Some sites are inactive/benign. Some (orange) are active and fear-inducing
• There’s a binary action agents can take: destroy some yellow sites
• The stimulus (ambush) landscape, and other agents, produces a disposition

to take the binary action

• Affect: Agents fear-condition on local stimuli: NOT conscious decision-making/choice.
• Passion
• Bounded rationality: Local sample relative frequency
• Reason
• The Sum = Solo Disposition (propensity to perform the act alone).
• Social animals: Add others’ weighted Solo Dispositions
• Weights are endogenous (minimizing parameters)
• If that Total Disposition exceeds threshold, take the action.
• Destroy, or flee, or refuse vaccine, or dump assets, or find guilty, binge on oreos…

​

Computational Parables : Slaughter of Innocents� Vision Von Neumann � Agent #0 fixed in SW: zero direct stimulus� Others in NE: stimulus, violent action � By dispositional contagion, Agent 0 acts.

One agent fixed in the

Peaceful south

Parable 1: Agent_Zero Joins�Without Direct Stimulus �(eye candy runs are just sample paths, of course)

​

​

Since no stimulus

within sensory radius.

Would not act alone

To understand why that happens, we need a quick look at the overall set-up.

Action , Threshold

• Binary Action
• Flee snake or don’t
• Raid icebox or don’t
• Join lynching or don’t
• Refuse vaccine or don’t
• Dump stock or don’t
• Wipe out village or don’t
• “Behavior” will mean a binary action.
• Nonnegative Real Threshold

Solo Disposition to Act

• Agents endowed with Affective V(t) and Deliberative P(t) real-valued functions bounded to [0,1] defined on a stochastic stimulus space.
• Their Solo disposition is, for the moment, as simple as possible, the sum:

​

​

​

​

​

But Socially Connected

• Agents also carry weights (unconsciously we presume).
• We therefore define the Total Disposition to Act as*

​

​

​

​

​

*self-weights assumed to be one, but can relax (low self-esteem agents).

Action Rule

����Dispositional Contagion, �Not Imitation of Behavior

• Nobody else’s observable action appears in this equation.
• Hence, the mechanism of action cannot be imitation of behavior, because the binary acts of others are not registered in this calculation.
• So we are suspending a “monkey-see/monkey-do” assumption central to much literature on social transmission.
• Obvious problem with imitation of observable action: no mechanism for the first actor. Nobody to imitate.
• (Noise is cheating…not a mechanism)

​

Further Under The Hood: Parsimony�

Agent Proper is Simple

• Three rules:
• Update affect
• Update probability
• Update disposition
• Act if disposition exceeds threshold
• Three Parameters
• The fear learning rate
• The memory window
• The threshold (equal across agents)
• Weights are endogenous, eliminating 2n parameters.

Further required apparatus…

• The stimulus landscape requires further parameters and code, initial conditions for everything, etc.
• But the agent alone is meant to be simple and transparent.

​

The Subtitle of Agent_Zero

• Toward Neurocognitive Foundations for Generative Social Science
• Talked about Generative Social Science
• What’s this neurocognitive business?

Fear Instantiation�

• Fear acquisition
• Fear extinction

​

​

Some neuroscience…

Amygdala Circuit

Amygdala Areas: Various Stains

Don’t Care Where…Care that it’s Innate, Automatic, Fast,�Inaccessible to Deliberation

Also equipped with an associative machinery.

“Neurons that fire together wire together.” Donald Hebb (1949)

Associative Fear Conditioning: �Acquisition Phase

​

US: Shock cuff

​

UR: Amygdala activation

​

CS: Blue Light (neutral)

​

CS-US Pairing Trials

Light…Shock

Light…Shock

Light…Shock

Light alone ………….🡪

Simple Elegant Model of Associative Learning �Rescorla-Wagner Model (1972)�

Learning rate is Surprise times Salience

Associative gain requires both.

(typically 1) is max associative strength.

Learning is fast when the stimulus is shocking,

but we become inured and learning slows.

Is fear contagious?

Observational Fear Conditioning

• Shown earlier : Fear-Conditioned human amygdala fMRI
• US: Shock cuff
• UR: Amygdala activation
• CS: Blue Light (neutral)
• CS-US PairingTrials
• Light…Shock
• Light…Shock
• Light…Shock
• Light alone ………… 🡪

​

​

Olsson, A., Nearing, K.I. and Phelps, E.A., 2007. Learning fears by observing others: the neural systems of social fear transmission. Social cognitive and affective neuroscience, 2(1), pp.3-11.

Is Fear Contagious?

• Top Panel (a), fMRI of subject above
• True Subject: Bottom Panel (b), fMRI of observer.
• Watches the blue-shock pairings
• Then is shown blue light alone…
• Same fMRI as if conditioned!
• Advantage clear
• I learn to fear the fire by watching you get burned
• But double-edged! Rapid non-conscious acquisition and transmission of baseless fears/biases.
• Fear can fade/decay if stimulus stops

​

​

​

​

​

​

​

​

​

Perils of Fitness

• “Survival circuits” (LeDoux 2012) conserved across vertebrate evolution.
• Epstein (2013) “Pleistocene man never encountered a BMW, but we freeze when a car whips around the corner at us, just as he froze when huge animals charged suddenly from the tall brush. We are harnessing the same innate fear-acquisition capacity—the same innate neurochemical computing architecture. Miraculously, synaptic plasticity permits us to adapt the evolved machinery to encode novel threats.”
• Invaluable but very dangerous…double-edged

Surprise + Salience 🡪 Strong Conditioning

Surprise + Salience 🡪 Strong Conditioning

Also Over-General and Persistent

Should stay afraid of hippos.

Affect can remain above the threshold long after actual stimulus has stopped.

If stimulus stops at t, extinction may be far off. Extreme case is PTSD.

​

How is this treated? In the learning phase, we had λ=1, maximum associative strength.

​

For extinction, we set this to zero.

​

Extinction in Rescorla-Wagner Model (1972)�

= 0 produces exponential decay

Full Affective Trajectory with Extinction�

Rats, predatory threat

We do not fear what the rat fears, but we fear how the rat fears.

With t* the time at which trials cease, the full solution is then

​

Ingredient 1: Emotion

• Introduce a generalized version of the classic (1972) Rescorla-Wagner model and emotional contagion through weights (endogenous functions of affect in book).

Reason may be “a slave to the passions,” a la Hume� but once in a great while, it happens…however badly!

• Typically we have incomplete and imperfect information
• Make systematically erroneous appraisals of it.
• Robustly documented errors:
• Framing effects (medical decisions)
• Endowment effects (loss aversion)
• Representativeness heuristic
• Local sample represents population
• Base rate neglect
• Confuse P(+|sick) with P(sick|+)
• Anchoring on what you hear first
• 2345678 < 8765432
• Agent_Zero (local relative frequentist) exhibits several.

​

​

To Make Matters Worse…

• Agents driven by strong (unconscious) emotions (like fear), doing bad statistics on incomplete and biased data, also influence one another.
• Conformist pressures can then produce widespread convergence on counter-productive behavior.
• Conformity effects are documented in many spheres (since Asch 1958).
• Again, a neural basis?

Yes: Nonconformity Hurts!

• Kross et al (PNAS 2011) “…when rejection is powerfully elicited…areas that support the sensory components of physical pain (secondary somatosensory cortex; dorsal posterior insula) become active.”
• Illustrated in fMRIs below.

​

​

​

Neural Drivers of Conformity

Neural overlap between social rejection and physical pain.

​

Bar graph: no statistically significant difference between (βs of) rejection and physical pain. Positive predictive value = 88%.

​

​

​

​

​

Source: Kross, E., Berman, M. G., Mischel, W., Smith, E. E., & Wager, T. D. (2011). Social rejection shares somatosensory representations with physical pain. Proceedings of the National Academy of Sciences of the United States of America, 108(15), 6270–6275.

Conform Because Rejection Hurts.

• As they write, “These results give new meaning to the idea that rejection ‘hurts’…rejection and physical pain are similar not only in that they are distressing—they share a common somatosensory representation as well.”
• We give others weight…so

Ingredient 3: Network Weights

• Agents experience a weighted sum of the affective and deliberative states of others
• As discussed, weights are actually endogenous in model—strength-scaled affective homophily generates networks…more on this below.

​

Given these components…

• Logic of the Model:
• Stimulus
• Disposition
• Threshold
• Action
• This feeds back to alter the stimulus pattern

Landscape and Trials:�Agent_0 Fixed and Mobile Rovers

Agents directly condition on orange trials and compute RF w/in vision.

Then a weighted sum over network. If D>τ, destroy all sites w/in

damage radius

​

Parable 1: Slaughter of Innocents� Agent 0 fixed, zero direct stimulus� Mobile rovers transmit retaliatory disposition� Vision Von Neumann…..Agent 0 massacres village

Parable 1: Agent_Zero Joins�Without Direct Stimulus

​

​

V=P=0, since no stimulus

within sensory radius

​

Solo disposition = 0

​

Eye candy is one sample path. Turn off and build statistical portrait.

La Condition Humaine

• Why?

​

​

• You take action in group (since ) that you would not take alone

(since ).

​

• Indeed, you may be the only agent with this ordering. In that case:

​

• Despite being negatively disposed* you act first!

​

​

*

Parable 2: Agent Zero Initiates

• Again, no direct stimulus

​

• He goes first!

​

• Not imitation of behavior

Core Parable: Agent_Zero Goes First Without Stimulus

Leadership or Susceptibility?

• Not behavioral imitation.
• If 1^st, nobody to imitate
• Leader, or just most susceptible to D-contagion?
• Tolstoy’s answer: ‘A king is history’s slave, performing for the swarm life.’ (War and Peace, 1896)

Network Extensions

• Scale-Up to large numbers
• Permit arbitrary exogenous network topologies (endogenous below)
• The most arresting Parable is the first actor:
• An agent that has no aversive stimulus and would not act alone leads the lynch mob, by dispositional contagion.
• How robust is that?
• For arbitrary numbers of northerners (with stimulus) and southerners (without) and arbitrary network topologies, it is mathematically formidable.
• Research underway with Dr. Jeewoen Shin.
• Preliminary computational experiments very interesting.

Arbitrary n with uniform or exponential

Explorations

Can scale up and stipulate fixed

network structures and explore

dynamics computationally

​

Exponential degree

distribution (λ=5)

​

Turn off all the movies,

assume distributions

and prove some theorems

on core phenomena:

[1] waiting time to first actor,

[2] probability of universal self betrayal.

Large-Scale Activation�without direct stimulus�by Dispositional Contagion

Large-Scale Activation without Direct Stimulus

Fight vs. Flight

• Fight

Flight

​

• Katrina Evacuees
• Syrian Refugees
• Capital/Portfolio Flight

​

• Recalcitrant agents “dragged out”

by others.

Would not flee her portfolio alone

In all of this, Networks are Implicated

• How do network weights change?

​

​

• Why do networks happen?

Endogenous Weight Change by Affective Homophily �(so weights are not parameters)

• Affective homophily. Affects changing. So try:
• Problem: equals zero when identical; want 1.0 when equal.
• OK, so as homophily, use:
• Problematic as a weight: nudniks (v=0) same strength as crusaders (v=1). So, scale by total strength

​

Lazer, David. "The co‐evolution of individual and network." Journal of Mathematical Sociology 25.1 (2001): 69-108.

Lazer, David, et al. "The coevolution of networks and political attitudes." Political Communication 27.3 (2010): 248-274.

Endogenous Weights:

• Affective homophily strengthens connection

​

• And this can matter immensely…

Grow The Arab Spring� Case 1: No Communication

Instances of regime corruption

(abduction, torture, theft, civil liberties)

​

​

Produce profound grievance

​

​

Weights clamped at zero by Big Brother.

​

​

In isolation, no action.

​

​

Arab Spring (Jasmine Revolutions)� Case 2: Communication🡪Dispositional Amplification🡪Overthrow

Revolt of the Swarm�

• Leaderless Revolutions
• No Mao, Lenin

​

​

• Similarly in Juries

​

Jury Dynamics: �12 Angry Agent_Zeros

• Pre-Trial: General landscape of stimuli about OJ’s guilt. Initial dispositions to convict are formed. Jurors strangers. All weights off.
• Trial: Competing stimuli (Prosecution and Defense). Dispositions are updated. Jurors do not communicate. Weights still off.
• Sequestration: Now homophily dynamics and network effects operate strongly.
• Agents convict in group when they would acquit alone:

​

​

Three-Phased Trial� Pre-trial Courtroom Jury Chamber

Pre-trial: S1>0, ω=0 Courtroom: S2>0, ω=0 Jury Phase: S3=0, ω>0

Jury Trial

Weights Jump in Jury Chamber.�Drive Dispositions to Convict�

Universal Self-Betrayal

No jurors would have convicted before the jury phase, but they

are unanimous in rendering a guilty verdict, having interacted directly.

​

​

​

“The overall picture of Homo sapiens reflected in these interpretations of

Agent_Zero is unsettling: Here we have a creature evolved (that is, selected)

for high susceptibility to unconscious fear conditioning. Fear (conscious

or otherwise) can be acquired rapidly through direct exposure or indirectly, through

fearful others. This primal emotion is moderated by a more

recently evolved deliberative module, which, at best, operates suboptimally

on incomplete data, and whose risk appraisals are normally biased further

by affect itself. Both affective and cognitive modules, moreover, are powerfully

influenced by the dispositions of similar—equally limited and unconsciously

driven—agents. Is it any wonder that collectivities of interacting

agents of this type—the Agent_Zero type—can exhibit mass violence, dysfunctional

health behaviors, and financial panic?” (Epstein, 2013)

Unsettling Picture

​

Extensions

Large NSF Proposal to Scale Up Dramatically

Populate National and Global Models with Cognitive Agents.

Parker, J. and Epstein, J.M., 2011. A distributed platform for global-scale agent-based models

of disease transmission. ACM Transactions on Modeling and Computer Simulation (TOMACS), 22(1), p.2.

Base Case Run: US National Model

Policy Applications: What is the effect if…

• We impose quarantines or social distancing measures?
• We close schools?
• We restrict entry into the country?
• When there is a vaccine, who should get it first: the elderly who are at greatest risk of death if infected, or young people who do most of the spreading?
• These are among the many goal of modeling we discussed earlier.
• They also arise in global pandemics, like the 1918 Spanish Flu and the COVID-19 pandemic. Here, we need an even larger-scale model…

Next Scale Up: Earth!� Global-Scale Agent Model� 6.5 Billion Agents

Epstein, Nature, 2009.

Parker and Epstein, TOMACS, 2011

​

​

​

Behavior and Mystery of Multiple Waves

Behavior and Multiple Waves:� “Spanish” Flu

COVID-19 US Data

Coupled Contagion Models and Waves

• Coupled Contagion models are producing multiple waves, by including human fears.
• Epstein et al,. “Coupled Contagion Dynamics of Fear and Disease” (PloS_ONE, 2008)
• Epstein et al,. “Triple Contagion: a Two Fears Epidemic Model” (Royal Society Interface, 2021)
• Cycles of behavior and disease dynamics.
• In the first paper, a single fear, of disease, Fd.

New cases↑⇒ Fd↑⇒ Contacts↓⇒

New cases↓⇒ Fd↓⇒ Contacts↑⇒

New cases↑… multiple waves, as in 1918 and COVID-19 pre-vaccine

• In the second, two fears, of disease Fd and of the control, vaccine Fv.

New cases↑⇒ Fd > Fv ⇒ Vaccine↑⇒

New cases↓⇒ Fd < Fv ⇒ Vaccine↓⇒

New cases↑… multiple waves, as in smallpox and COVID-19 with vaccine, continuing w/ delta variant

• In the Triple Contagion model, the disease and the two fears all interact to reveal new behavioral wave mechanisms.

​

​

Disease and Fear of Disease

Second Peak > First Peak

Vaccine Refusal

Another behavior that can profoundly affect the course of an epidemic—possibly undermining herd immunity—is the refusal of vaccine.

The World Health Organization includes vaccine refusal as one of the top ten threats to global health.

It is responsible for the resurgence of vaccine-preventable diseases including measles and, in some parts of the world, even polio.

Central in COVID today!!

People may be afraid of the disease but may be even more afraid of the vaccine.

Disease and contagious fear of the vaccine

High adverse event rate σ

increases vaccine-fear and

refusal, allowing resurgence, which can be higher than the first peak.

Other Extensions

• Addict_Zero with Erez Hatna and Jeewoen Shin
• Atheist_Zero with Erez Hatna and Annetta Burger

​

• Many other activities at our Lab…please check them out!

• Thank you!

​

• Please feel free to follow up: je65@nyu.edu